Ayati, S. A., Park, J. H., Cai, Y., & Botacin, M. 2025. Making Acoustic Side-Channel Attacks on Noisy Keyboards Viable with LLM-Assisted Spectrograms’ Typo Correction. The 19th USENIX WOOT Conference on Offensive Technologies (WOOT ’25).
@inproceedings{ayati2025echo,
author = {Ayati, Seyyed Ali and Park, Jin Hyun and Cai, Yichen and Botacin, Marcus},
title = {Making Acoustic Side-Channel Attacks on Noisy Keyboards Viable with LLM-Assisted Spectrograms' Typo Correction},
booktitle = {The 19th USENIX WOOT Conference on Offensive Technologies (WOOT '25)},
year = {2025},
url = {https://www.usenix.org/conference/woot25/presentation/ayati},
doi = {}
}
Acoustic side-channel attacks (ASCAs) have been shown to successfully recover keystrokes from recordings of mechanical keyboard clicks. However, previous work assume controlled conditions — quiet environments and high-quality recordings — that are rarely met in real-world scenarios. In this paper, we show that LLM-assisted post-processing of spectrogram transcripts can make ASCAs viable under noisy, unconstrained conditions, achieving state-of-the-art accuracy on consumer-grade recordings.